Realtime gesture recognition under the multi-layered parallel recognition framework of QVIPS

The importance of gesture protocols has caught little attention in the field of gesture-related computer vision research. The gesture protocols are the consensus on gestures, which are formed through face-to-face communication in our daily lives. Traditional gesture recognition algorithms unilaterally presupposed the gesture protocol without the consensus from users. As a consequence, users were directed to follow the rigidly prescribed gestures unintentionally. Standing firm on the belief that the gestures are not substitute for keyboards and mice, we propose a flexible gesture recognition framework which can adapt to any kinds of gestures presented by the user him/herself. To focus on a process in which the gesture protocols are formed, the Quadruple Visual Interest Point Strategy (QVIPS) is newly introduced. QVIPS enables the system to observe a gesture from multilateral perspectives and to recognize the position, posture and motion information of a user under a unified framework. Inevitably QVIPS requires additional computational burden. We deal with this problem by the combinational use of a Gaussian Density Feature (GDF) and fast Fourier transform (FFT). The proposed system performs recognition in real-time without any special purpose hardware.